Comparative benchmark of machine learning models for predicting perovskite solar cell performance

Abstract

This study presents a systematic comparative evaluation of five machine learning models, namely Random Forest (RF), Extreme Gradient Boosting (XGB), Artificial Neural Network (ANN), Multilayer Perceptron (MLP), and Support Vector Machine (SVM), for predicting the electrical performance of perovskite solar cells (PSCs). The study is based on a dataset of 18,570 device configurations generated using SCAPS-1D simulations, covering a wide range of materials, geometries, and optoelectronic parameters. Among the evaluated models, tree-based models show the best performance, with XGB attaining the highest R 2 for open-circuit voltage and short-circuit current density (0.9774 and 0.9815, respectively). At the same time, RF minimizes absolute errors for the fill factor (FF), with root mean squared error (RMSE) of 0.8978 and mean absolute error (MAE) of 0.2909, and power conversion efficiency (PCE) with RMSE of 0.6307 and MAE of 0.2338. Permutation-importance analysis, performed for both RF and XGB, identified the perovskite bandgap, defect density, and absorber thickness as recurrent key variables governing the photovoltaic response. These results are consistent with bandgap-controlled absorption and voltage generation, defect-assisted non-radiative recombination, and thickness-dependent carrier generation and collection. To evaluate the model generalization, RF predictions were validated against independent results reported in experimental and numerical studies, showing moderate deviations and consistent performance trends. Thus, the results reported in this work establish a reproducible machine learning workflow that bridges numerical simulation and physical interpretation, while also supporting its use as a surrogate-modeling strategy for the preliminary screening of PSC architectures.